The AI Triad and What It Means for National Security Strategy

AUGUST 2020 AUTHOR The AI Triad and Ben Buchanan What It Means for National Security Strategy Established in January 2019, the Center for Security and Emerging Technology (CSET) at Georgetown’s Walsh School of Foreign Service is a research organization focused on studying the security impacts of emerging technologies, supporting academic work in security and technology studies, and delivering nonpartisan analysis to the policy community. CSET aims to prepare a generation of policymakers, analysts, and diplomats to address the challenges and opportunities of emerging technologies. During its first two years, CSET will focus on the effects of progress in artificial intelligence and advanced computing. CSET.GEORGETOWN.EDU | [email protected] 2 Center for Security and Emerging Technology AUGUST 2020 The AI Triad and What It Means for National Security Strategy AUTHOR Ben Buchanan ACKNOWLEDGMENTS The author would like to thank John Bansemer, Beba Cibralic, Tarun Chhabra, Teddy Collins, Andrew Imbrie, Igor Mikolic-Torreira, Michael Sulmeyer, Danielle Tarraf, Lynne Weil, and Alexandra Vreeman for their comments on an earlier draft of this paper. PRINT AND ELECTRONIC DISTRIBUTION RIGHTS © 2020 by the Center for Security and Emerging Technology. This work is licensed under a Creative Commons Attribution- NonCommercial 4.0 International License. To view a copy of this license, visit: https://creativecommons.org/licenses/by-nc/4.0/. Document Identifier: doi: 10.51593/20200021 Cover photo: paul/AdobeStock Contents EXECUTIVE SUMMARY III INTRODUCTION V 1 | THE AI TRIAD 1 2 | WHAT THE AI TRIAD MEANS FOR POLICYMAKERS 11 CONCLUSION 15 ENDNOTES 17 Center for Security and Emerging Technology i iv Center for Security and Emerging Technology Executive Summary single sentence can summarize the complexities of modern artificial intelligence: Machine learning systems use computing A power to execute algorithms that learn from data. Everything national security policymakers truly need to know about a technology that seems simultaneously trendy, powerful, and mysterious is captured in those 13 words. They specify a paradigm for modern AI—machine learning—in which machines draw their own insights from data, unlike the human-driven expert systems of the past. The same sentence also introduces the AI triad of algorithms, data, and computing power. Each element is vital to the power of machine learning systems, though their relative priority changes based on technological developments. Algorithms govern how machine learning systems process information and make decisions. Three main classes of algorithms are common today: supervised learning, which draws insights from struc- tured data sets; unsupervised learning, which excels at finding structure or clusters in unorganized data sets; and reinforcement learning, which builds up a machine learning system’s capability through trial and error. Usual- ly, these algorithms run on neural networks, a type of computer program architecture. Neural networks provide enormous flexibility and power, but come with their own tradeoffs—chiefly, the lack of transparency behind their reasoning. Data is often, though not always, how machine learning systems learn about the world. If a fact is not in the data provided to the machine, the system will never learn it, especially in the case of supervised learning. Acquiring larger and more representative datasets can further empower Center for Security and Emerging Technology iii machine learning systems. Without such datasets, bias can creep into systems, caus- ing them to perform poorly in hard-to-detect ways. Often overlooked, computing power is increasingly essential as algorithms grow more complex and datasets larger. The past eight years have seen a revolu- tion in the amount of computing power applied to cutting-edge machine learning, expanding by a factor of several hundred thousand. Computing power increasingly affects the performance of machine learning systems and presents a significant cost during system development. Each part of the triad offers its own policy levers. Algorithmic progress depends on a nation acquiring and developing talented machine learning researchers. Larger and better datasets require tricky policy choices involving bias, privacy, and cybersecurity. Computing power can provide a point of leverage for export controls in foreign policy, as well as a bottleneck for AI research at home. In order to judi- ciously wield the levers available in AI policy, policymakers must first understand the technology itself and how it will reshape national security. The concept of the AI triad is one framework for doing so. iv Center for Security and Emerging Technology Introduction single sentence can summarize the complexities of modern artificial intelligence: Machine learning systems use computing power A to execute algorithms that learn from data. Everything policymakers need to know about a technology that seems simultaneously trendy, powerful, and mysterious is captured in those 13 words. AI matters. At home, it is already fundamental to everyday life, voiced by Alexa and Siri and tucked inside smartwatches and phones. In science, it contributes to major breakthroughs, from diagnosing disease to aiding drug discovery to modeling the climate. In business, it shapes the economic competitiveness of nations and alters how trillions of dollars pulse through global markets. In national security, it bolsters logistics and intelligence analysis and—with visions of lethal autonomous weapons, drone warfare, and self-guided cyberattacks—seems poised to do much more. The breadth of the technology is why the single-sentence articulation of AI is so important, and why the concepts alluded to within it matter so much. If policymakers do not understand AI, they will be a passive audi- ence to technological pioneers charging onward, slow to recognize what AI can do for the issues they care about. Maybe worse, policymakers who do not understand AI will fail to recognize what the technology cannot yet do, despite its hype. They will ignore AI’s current structural flaws, such as a lack of transparency and the potential for insidious bias—challenges that must be mitigated with both technical and policy solutions. The concise definition of AI first specifies a paradigm for modern AI: machine learning. Machine learning stands in direct contrast to the previ- ous era’s dominant paradigm, expert systems, which focused on formally Center for Security and Emerging Technology v encoding human knowledge in a way a computer could process. For example, IBM’s computer program DeepBlue drew heavily on inputs provided in advance by chess grandmasters to beat world chess champion Garry Kasparov in 1997. In machine learning, AI developers reject direct instructions in favor of a system that can learn on its own. This paper focuses on the dominant paradigm of machine learning known as deep learning, explained in more detail in the next section. Three components make deep learning happen: data, algorithms, and computing power. Together, I call these components the AI triad. Computer scientists have long used this tripartite division to study machine learning, and I argue that it offers a framework for understanding the power of machine learning and what it means for policy. The elements of the AI triad work in combination to achieve stunning results. For example, OpenAI, a leading AI research lab, trained a text generation system, known as GPT-3, to write whole paragraphs in response to a given prompt and to perform other linguistic tasks. The engineers assembled almost a trillion words, which they filtered down to around 540GB of human writing. They then devised a clever algorithm with around 175 billion learned parameters that could predict which word would come next in a sentence based on patterns in the collected data; in essence, the algorithm learned to imitate the writing it saw. The engineers set GPT-3 loose on high-performance computers for days, performing quadrillions of calculations as it refined its own capacity for mimicking human language. GPT-3’s combination of data, algorithms, and computing power produced a breakthrough. The system wrote some text that 88 percent of readers thought was convincingly human. Perhaps most impressive was GPT-3’s ability to mimic its prompt, from continuing a news report to writing the next stanza of a poem in the style of a particular poet.1 GPT-3 is one in a long line of machine learning breakthroughs. The drumbeat of advancement and machine learning’s relevance to national security show no signs of diminishing. Policymakers need a deeper understanding of machine learning’s three components and why they matter so much. vi Center for Security and Emerging Technology 1 The AI Triad ALGORITHMS An algorithm is a series of instructions for processing information. Schoolchildren, for example, are taught the algorithm PEMDAS for the order of operations in solving math problems: parentheses, exponents, multiplication/division, addition/subtraction. When given information— such as the problem 7+5(1+3)—the algorithm tells them explicitly how to process it: first add one and three in the parentheses, then take the resulting four and multiply it by the five next to it, and finally add the resulting 20 to the seven. Another algorithm, one that works from left to right and ignores order of operations, processes the information differently—adding seven to five, multiplying by one, and then adding three. It yields a different (and incor- rect) answer. In the same way, computer programmers issue direct instructions—algorithms—to their systems on how to process information. These algorithms often contain conditional logic—if this, then that—specifying that a program should do one thing with one set of information but a different thing with a different set of information. For decades, engineers built AI programs with a similar kind of design. As mentioned, DeepBlue’s defeat of chess champion Garry Kasparov was enabled by a detailed series of direct instructions culled from grandmasters that guided that program’s play in the tournament match.

Load more